Orchestrate Codex

Run multiple codex CLI workers under one manifest, one Monitor, one policy. The skill detects mode → seeds a manifest → spawns a detached runner → emits a Monitor hint → exits. Codex workers run in the background; the agent stays in the conversation; the manifest is the source of truth.

This skill is for codex-only orchestration. PRs go to ask-review. Bot-comment triage goes to evaluate-code-review. Cross-LLM batch fanout is out of scope.

Trigger Boundary

Use this skill when the task matches one of:

• parallel codex agents on ≥2 discrete coding tasks across separate worktrees, run as one fleet under one manifest
• one prompt template fanned out across N inputs (URLs, IDs, file paths) with one answer file per input
• one substantial codex mission the user wants to watch live via streamed JSONL events into Monitor
• a list of branches that need iterative codex exec review rounds with classifier-driven converged/blocked decisions
• a prior orchestrate-codex run that did not finish — resuming failed entries, never-started entries, or all non-done from the existing manifest
• any codex fleet that requires shared sandbox/model/effort policy, bounded concurrency, atomic manifest writes, and one Monitor surface

Do NOT use this skill when: